Retrieval that makes a model
answer from your data.

What RAG is, and what it fixes

A language model on its own answers from what it learned in training. It does not know your contracts, your ledger, your policies or last quarter’s filings, and when asked it will produce something fluent and unsupported.

Retrieval-Augmented Generation puts a search step in front of the model. The question is used to retrieve the relevant passages from your own corpus; those passages are given to the model; the answer is generated from them and cites them. The model stops being the source of truth and becomes the thing that reads and summarises it.

The pipeline, stage by stage

Ingestion

Documents, records, wikis and database rows pulled in from where they already live, with their permissions preserved rather than flattened.

Chunking

Splitting a document so a retrieved passage is self-contained. This is where most enterprise RAG systems are won or lost — chunk badly and retrieval returns fragments that read as nonsense.

Embeddings

Passages encoded into vectors by an embedding model that runs inside your environment, so the corpus is never sent out to be indexed.

Vector search

Retrieval over those vectors, usually hybrid — dense vectors for meaning, keyword search for exact terms like a policy number or a counterparty name.

Reranking

A second pass that reorders candidates by actual relevance. Retrieval gets you close; reranking is what makes the top three worth reading.

Grounded generation

The answer is produced from the retrieved passages and cites them, so a reader can check it.

What enterprise RAG is, and how it differs

Enterprise RAG is retrieval-augmented generation built against a real organisation’s corpus rather than a clean test set: retrieval filtered by the asking user’s permissions, source documents that contradict each other and were never meant to be parsed, and an index that has to stay correct while the corpus changes. Document intelligence is the layer beneath it — extracting structure, fields and meaning from documents written for people instead of for machines.

A demo RAG system indexes a folder of PDFs. An enterprise one has to respect who is allowed to see what, handle documents that contradict each other, cope with formats that were never meant to be parsed, and stay correct as the corpus changes underneath it.

Permissions are the part most often skipped. If retrieval ignores access control, the system will cheerfully quote a document to somebody who was never allowed to open it. Retrieval has to be filtered by the asking user’s entitlements, not filtered afterwards.

Private RAG

Because the corpus is the sensitive asset, retrieval is the part of an AI system that most needs to stay inside. Embeddings are derived from your documents and can leak their content; an index held by a third party is your data in another form.

We build RAG to run entirely within your perimeter — the embedding model, the vector store, the reranker and the generating model all inside, on-premises, in a private cloud tenancy or air-gapped.